(6694 words) Dancing with Pixies: Strong Artificial Intelligence and Panpsychism Mark Bishop In 1994 John Searle stated (Searle 1994, pp.11-12) that the Chinese Room Argument (CRA) is an attempt to prove the truth of the premise: 1: Syntax is not sufficient for semantics which, together with the following: 2: Programs are formal, 3: Minds have content led him to the conclusion that ‘programs are not minds’ and hence that computationalism, the idea that the essence of thinking lies in computational processes and that such processes thereby underlie and explain conscious thinking, is false. The argument presented in this paper is not a direct attack or defence of the CRA, but relates to the premise at its heart, that syntax is not sufficient for semantics, via the closely associated propositions that semantics is not intrinsic to syntax and that syntax is not intrinsic to physics.1 However, in contrast to the CRA’s critique of the link between syntax and semantics, this 1 See Searle (1990, 1992) for related discussion. 1 paper will explore the associated link between syntax and physics. The main argument presented here is not significantly original – it is a simple reflection upon that originally given by Hilary Putnam (Putnam 1988) and criticised by David Chalmers and others.2 In what follows, instead of seeking to justify Putnam’s claim that, “every open system implements every Finite State Automaton (FSA)”, and hence that psychological states of the brain cannot be functional states of a computer, I will seek to establish the weaker result that, over a finite time window every open system implements the trace of a particular FSA Q, as it executes program (p) on input (x). That this result leads to panpsychism is clear as, equating Q (p, x) to a specific Strong AI program that is claimed to instantiate phenomenal states as it executes, and following Putnam’s procedure, identical computational (and ex hypothesi phenomenal) states (ubiquitous little ‘pixies’) can be found in every open physical system. The route-map for this endeavour is as follows. In the first part of the paper I delineate the boundaries of the CRA to explicitly target all attempts at machine understanding – not just the script-based methods of Schank and Abelson (Schank & Abelson 1977). Secondly I introduce Discrete State Machines, DSMs, and show how, with input to them defined, their behaviour is described by a simple unbranching sequence of state transitions analogous to that of an inputless FSA. Then I review Putnam’s 1988 argument that purports to show how every open physical system implements every inputless FSA. This argument is subsequently applied to a robotic system that is claimed to instantiate genuine phenomenal states as it operates. Thus, unlike the CRA, which primarily concerns the ability of a suitably programmed computer to understand, this paper outlines a reductio- 2 See Chalmers (1994, 1996a, 1996b) and also the special issue, What is Computation? of Minds and Machines, (vol.4, no.4, November 1994). 2 style argument against the notion that a suitably programmed computer qua performing computation can ever instantiate genuine phenomenal states. I conclude the paper with a discussion of three interesting objections to this thesis. The Chinese Room The twenty years since its inception have seen many reactions to the Chinese Room Argument from both the philosophical and cognitive science communities. Comment in this volume ranges from Bringsjord, who asserts the CRA to be “arguably the 20th century’s greatest philosophical polarizer”, to Rey who claims that in his definition of Strong AI, Searle, “burdens the [Computational Representational Theory of Thought (Strong AI)] project with extraneous claims which any serious defender of it should reject”. Yet the CRA is not a critique of AI per se – indeed it is explicit in ‘Minds, Brains, and Programs’, as in other of his expositions, that Searle believes that there is no barrier in principle to the notion that a machine can think and understand. The CRA is primarily a critique of computationalism, according to which a machine could have genuine mental states (e.g. genuinely understand Chinese) purely in virtue of its carrying out a series of computations. In the CRA Searle presented a rebuttal of the then computationalist orthodoxy that viewed cognition and intelligence as nothing more than symbol manipulation and search.3 Following work on the automatic analysis of simple stories, a cultural context emerged within the AI community that appeared comfortable with the notion that computers were able to ‘understand’ such stories, a concept which can be traced back to the publication of Alan Turing’s seminal paper ‘Computing Machinery and Intelligence’ (Turing 1950). 3 E.g. Newell & Simon 1976. 3 For Turing, emerging from the fading backdrop of Logical Positivism and the Vienna Circle, conventional questions concerning ‘machine thinking’ were too imprecise to be answered scientifically and needed to be replaced by a question that could be unambiguously expressed in scientific language. In considering the metaphysical question, ‘Can a machine think?’, Turing arrived at the other, distinctly empirical, question of, whether, in remote interaction via teletype with both a computer and a human, a human could identify which was which as accurately as by chance. If so, the computer is said to have passed the Turing Test. It is now more than fifty years since Turing published details of his test for machine intelligence and although the test has since been discredited by several commentators (e.g. Bringsjord 1992, Kelly 1993), the notion of a thinking machine continues to flourish. Indeed, the concept has become so ingrained in popular culture by science fiction books and movies that many consider it almost apostate to question it. And yet, given the poverty of current AI systems on relatively simple linguistic comprehension problems,4 it is hardly surprising that, when writing on the subject, a phrase from Hans Christian Andersen slipped into Roger Penrose’s mind (Penrose 1989). Nonetheless, throughout the 1970’s and early 1980’s, Searle and Hubert Dreyfus (Dreyfus 1972) remained isolated voices that surfaced above the hegemony of symbolically Strong AI. Still today, partly due to the (apparent) simplicity of its attack, the CRA is perhaps the best-known philosophical argument in this area. In the CRA Searle argues that understanding of a Chinese story can 4 As illustrated by the poor quality of the entrants to the annual Loebner prize (an award made to the program that can best maintain a believable dialogue with a human. See http://www.loebner.net/Prizef/loebner-prize.html). 4 never arise purely as a result of the state transforms caused by ‘following the instructions’ of any computer program. His original paper offers a first-person tale outlining how Searle could instantiate such a program, produce correct internal and external state transitions, pass a Turing Test for understanding Chinese, and yet still not understand a word of Chinese. However, in the twenty years since its publication, perhaps because of its ubiquity and the widespread background perception that, if it succeeds at all, its primarily target is Good Old-Fashioned AI (GOFAI), the focus of AI research has drifted into other areas: connectionism, evolutionary computing, embodied robotics, etc. Because such typically cybernetic5 approaches to AI are perceived to be the antithesis of formal, rule-based, script techniques, many working in these fields believe the CRA is not directed at them. Unfortunately it is, for Searle’s rule-book of instructions could be precisely those defining learning in a neural network, search in a genetic algorithm or even controlling the behaviour of a humanoid-style robot of the type beloved by Hollywood. But what does it mean to genuinely understand Chinese? That it is not simply a matter of acting in the behaviourally correct way is illustrated if we consider Wittgenstein’s illustration of the difference between following a rule and merely acting in accordance with it.6 Although rule-following requires regularity in behaviour, regularity alone is not enough. The movements of planets are correctly described by Kepler’s laws, but planets do not follow those laws in a way that constitutes rule-following behaviour. It is clear that the CRA employs a similar rhetorical device. It asks: ‘Does the appropriately programmed computer follow the rules of (i.e. 5 Cybernetic AI is characterised by emphasis on ‘sub-symbolic knowledge representation’ and a ‘bottom-up’ approach to problem solving. 6 Wittgenstein 1953, §§207-8, 232. 5 understand) Chinese when it generates ‘correct’ responses to questions asked about a story, or is it merely that its behaviour is correctly described by those rules?’. This difference between genuinely following a rule and merely acting in accordance with it seems to undermine Turing’s unashamedly behaviouristic test for machine intelligence. That the CRA addresses both phenomenal and intentional aspects of understanding and intelligence is clear from the introduction to Searle’s original paper, where we find Searle’s definition of Strong AI: But according to Strong AI the computer is not merely a tool in the study of the mind; rather the appropriately programmed computer really is a mind, in the sense that computers given the right programs can be literally said to understand and have other cognitive states. (Searle 1980, p.417 (p.67 in Boden)). An axial statement here is that, ‘the appropriately programmed computer really is a mind’. This, taken in conjunction with, ‘[the appropriately- programmed computer] can be literally said to understand’ and hence have associated ‘other cognitive states’, implies that the CRA also, at the very least, targets some aspects of machine consciousness – the phenomenal infrastructure that goes along with ‘really having a mind’.